Activities
Divisions
Performances
Activities
Divisions
Performances
Apple has at last introduced MM1, its combined AI model for producing text and images. The AI model is also capable of making predictions within context due to its extensive pre-training in multiple modes.
Following a period of gossip and conjecture about their forthcoming AI endeavors and multimodal AI models, Apple's research team has created a group of substantial multimodal language models named MM1. These models are capable of analyzing and producing both textual and visual information, as stated in a research paper revealed last week.
The research conducted at Apple's labs was focused on developing efficient and robust multimodal large language models (MLLMs) by meticulously examining and modifying different architectural elements, data sources, and training methods.
The study discovered that the quality of the image and the capability of the visual encoder greatly influenced the performance of the model, whereas the particular technique of merging visual and textual data was not as significant.
They also found that meticulously blending various types of data was vital. Few-shot learning was aided by intermixed image-text documents, while conventional captioned images enhanced zero-shot performance. Moreover, incorporating text-only data helped in preserving robust language comprehension abilities.
MM1 is capable of making predictions within a specific context due to its comprehensive multimodal pre-training. It has the ability to count items and adhere to bespoke formatting, draw references from image sections and execute OCR. Moreover, it can showcase general understanding and vocabulary related to everyday items, as well as carry out elementary mathematical functions.
Drawing from these observations, the group devised the MM1 model assortment, which varied from three billion to 30 billion parameters, encapsulating both dense and mixture-of-experts versions. Following the amplification of training, MM1 accomplished top-tier outcomes on diverse multimodal standards during the pre-training phase.
After additional fine-tuning using a carefully chosen dataset of 1 million examples, the ultimate MM1 models exhibited strong performance in 12 different multimodal tasks, including visual question answering and captioning. Importantly, MM1 was capable of handling multiple-image reasoning and few-shot learning, crucial skills made possible through the team's meticulous multimodal pre-training strategy.
This study further develops prior investigations into subjects such as CLIP for acquiring visual representations through natural language supervision, and autoregressive models such as GPT for generating text. Nonetheless, it stands as one of the pioneering extensive researches concentrating particularly on large-scale multimodal pre-training.
The researchers are optimistic that their findings will fast-track advancements, as it is rumored that Apple is negotiating to incorporate Google's Gemini generative AI models into future iPhone software.
Look for us on YouTube
Featured Programs
Connected Articles
Microsoft recruits Mustafa Suleyman, co-founder of DeepMind, to head their new consumer AI division
Samsung and Rebellions, South Korean chip producers, aim to surpass NVIDIA
The upcoming Apple Watch generation is expected to include a long-anticipated feature
NVIDIA unveils their new Blackwell B200 AI superchip, boasting a power 30 times greater than their existing top-tier H100
Microsoft brings in Mustafa Suleyman, co-founder of DeepMind, to spearhead their fresh consumer AI team
Samsung and Rebellions, chip makers from South Korea, are strategizing to dethrone NVIDIA
The next Apple Watch model is set to finally provide a feature that has been eagerly awaited
NVIDIA introduces their latest Blackwell B200 AI superchip, which they claim has 30 times the power of their present flagship model, the H100
Available on YouTube
Firstpost retains all rights, copyright @ 2024.
+ There are no comments
Add yours